github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-04-29 03:00:14 -04:00
Code Issues Packages Projects Releases Wiki Activity

use default dict for external_model_benchmark (#2592)

Browse Source 
* device default

* Device.DEFAULT

* half max for cuda

* CUDA_INCLUDE_PATH

* closer to working

* cuda fixups

* Update ops_cuda.py
This commit is contained in:
George Hotz
2023-12-03 15:25:43 -08:00
committed by GitHub
parent 550817389a
commit bbeba8ec85
7 changed files with 64 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
test/external/external_cl_half_max.py vendored Normal file
View File

@@ -0,0 +1,13 @@
from tinygrad.runtime.ops_gpu import CLDevice, CLProgram, compile_cl
if __name__ == "__main__":
dev = CLDevice()
lib = compile_cl("""
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
__kernel void test(__global half *out, __global half *a, __global half *b) {
int gid = get_global_id(0);
out[gid] = max(a[gid], b[gid]);
}
""")
prg = CLProgram(dev, "test", lib)

14
test/external/external_model_benchmark.py vendored
View File

@@ -43,7 +43,7 @@ def benchmark(mnm, nm, fxn):
#BASE = pathlib.Path(__file__).parents[2] / "weights" / "onnx"
BASE = pathlib.Path("/tmp/onnx")
def benchmark_model(m, validate_outs=False):
def benchmark_model(m, devices, validate_outs=False):
torch.manual_seed(1)
global open_csv, CSV
CSV = {"model": m}
@@ -61,7 +61,7 @@ def benchmark_model(m, validate_outs=False):
# print input names
if DEBUG >= 2: print([inp.name for inp in onnx_model.graph.input if inp.name not in excluded])
for device in ["METAL" if OSX else "GPU", "CLANG"]: # + (["CUDA"] if torch.cuda.is_available() else []):
for device in devices:
Device.DEFAULT = device
inputs = {k:Tensor(inp) for k,inp in np_inputs.items()}
tinygrad_model = get_run_onnx(onnx_model)
@@ -92,11 +92,14 @@ def benchmark_model(m, validate_outs=False):
provider = backend+"ExecutionProvider"
if provider not in ort.get_available_providers(): continue
ort_sess = ort.InferenceSession(str(fn), ort_options, [provider])
benchmark(m, f"onnxruntime_{backend.lower()}", lambda: ort_sess.run(output_names, np_inputs))
try:
benchmark(m, f"onnxruntime_{backend.lower()}", lambda: ort_sess.run(output_names, np_inputs))
except Exception as e: print(f"{m:16s}onnxruntime_{backend.lower()} {type(e).__name__:>25}")
del ort_sess
if validate_outs:
rtol, atol = 2e-3, 2e-3 # tolerance for fp16 models
if m == "openpilot" and 'CUDA' in devices: rtol, atol = 0.1, 0.1 # TODO: why is this broken?
inputs = {k:Tensor(inp) for k,inp in np_inputs.items()}
tinygrad_model = get_run_onnx(onnx_model)
tinygrad_out = tinygrad_model(inputs)
@@ -121,6 +124,7 @@ def assert_allclose(tiny_out:dict, onnx_out:dict, rtol=1e-5, atol=1e-5):
else: np.testing.assert_allclose(tiny_v.numpy(), onnx_v, rtol=rtol, atol=atol, err_msg=f"For tensor '{k}' in {tiny_out.keys()}")
if __name__ == "__main__":
if getenv("MODEL", "") != "": benchmark_model(getenv("MODEL", ""), True)
devices = [Device.DEFAULT] if getenv("NOCLANG") else [Device.DEFAULT, "CLANG"]
if getenv("MODEL", "") != "": benchmark_model(getenv("MODEL", ""), devices, True)
else:
for m in MODELS: benchmark_model(m, True)
for m in MODELS: benchmark_model(m, devices, True)